→
Decisions first
Models exist to support a decision. That decision and its cost of error get defined before optimizing any metric.
// about
A studio built around trustworthy ML.
MaineFrame Labs is the freelance practice of a machine learning engineer with a PhD and hands-on industry experience — someone who spends more time on data, uncertainty, and evaluation than on chasing leaderboard numbers.
Vedran Stipetic
// founder · PhD Computer Vision · MSc Applied Mathematics
I'm the person behind MaineFrame Labs. I help companies build and deploy production ML systems — from prototype to scale — combining deep learning with mathematical modeling. That pairing matters most for problems that need a principled handling of uncertainty and physical constraints, not just a model that fits the training set.
Before going independent, I built visual search and image-quality / image-ranking systems at Tripadvisor, working at 300M+ image scale. Earlier work includes real-time perception for autonomous vehicles and satellite-imagery ML pipelines, alongside published research in image restoration.
- Computer Vision
- Deep Learning
- Generative AI
- Data-Centric AI
- Edge Deployment
- Visual search & image quality — ranking systems at 300M+ image scale (Tripadvisor).
- Autonomous-vehicle perception — real-time, edge-deployed vision.
- Satellite imagery pipelines — detection & analysis under data scarcity.
- Published research — image restoration.
The approach
The studio is led by a PhD researcher with industry experience — a combination that's genuinely rare in freelance ML. Academic rigor (honest evaluation, calibration, careful baselines) paired with the pragmatism of shipping systems that have to work in production. That background is the difference between a model that benchmarks well and one you can actually trust.
Most ML projects don't fail at the model — they fail at the boundary between the model and the real world: messy data, drifting distributions, and predictions that look confident but shouldn't. MaineFrame Labs is built around that boundary.
Work here pairs strong modeling fundamentals with a data-centric mindset: active learning to spend labeling budget where it counts, and uncertainty quantification to know which predictions are safe to act on. The output is not just a model, but a reproducible, observable system you can trust in production.
Engagements are typically freelance and project-based, with the option to embed alongside an in-house team when that fits better.
Focus areas
- Computer vision — detection & segmentation, incl. satellite / remote-sensing imagery.
- Multimodal generative AI — building with VLMs and studying their limits.
- Active learning — efficient labeling loops for scarce-label domains.
- Uncertainty quantification — calibration, UQ, and knowing when not to trust a model.
- Evaluation & explainability — honest assessment of what models actually do.
Principles
How the work gets done
~
Uncertainty is a feature
A confident wrong prediction is worse than an honest "I don't know." Calibration and UQ are first-class deliverables.
∞
Reproducible by default
Pipelines, evaluation harnesses, and clear documentation — so the work survives after the engagement ends.